Core Concepts
Getting Into Claude Code — The Building Blocks of Your Personal AI Assistant
Core Concepts, part 2. What Claude Code is, what CLAUDE.md, skills, MCP, and subagents each do, plus usage concepts like sessions and Plan Mode — explained through analogies.
This is part 2 of the Core Concepts series. Part 1 covered the principles of LLMs and agents; this post covers the tool built on top of them — Claude Code. You can read this without part 1, but if "agent" and "harness" sound unfamiliar, start there.
What is Claude Code?
In one line:
Claude Code is not a chatbot — it's an agent. Not a tool you ask questions, but a tool you give work to.
In the language of part 1, Claude Code is a harness — the shell that runs an agent safely. It keeps the agent from running wild, makes it ask before risky operations, and confines its work to your folder. Claude Code is the harness Anthropic built.
The name says "Code," but it isn't just a coding tool. Simon Willison, a well-known developer in the AI space, put it this way:
"Claude Code is, with hindsight, poorly named. It's not purely a coding tool: it's a tool for general computer automation."
Email drafts, reports, data cleanup, research, file management — most of the repetitive work you do on a computer can be handed over.
To sum it up as an analogy: the LLM is a smart new hire's brain, the agent is the new hire actually showing up to work, and Claude Code is the office and the work rules that hire operates under.
Choosing a model: Haiku, Sonnet, Opus
Inside Claude Code, the /model command lets you swap brains.
| Model | Character | When to use it |
|---|---|---|
| Haiku | Fast and economical | Simple questions, translation, summaries |
| Sonnet | Balanced performance | Most day-to-day work |
| Opus | Deepest reasoning | Complex analysis, strategy, long documents |
If your usage allowance is generous, use Opus; if you want to conserve it, use Sonnet. That's all the rule you need.
The four core building blocks
Four things turn Claude Code into "your personal AI assistant." Each one you assemble makes the assistant know you better.
CLAUDE.md — the work manual you hand the AI
A file Claude reads automatically at the start of every session. It's like the onboarding manual a company gives a new hire. Instead of repeating "here's who I am, here's how I want things done" every time, write it once and Claude works with it always in mind.
# AI Assistant Manual
## Basics
- I run a company in [industry]
- Always respond in a professional tone
- Keep reports concise (one page max)
## Frequent tasks
- Client emails in a courteous tone
- Always present numbers in tables
## Don'ts
- No jargon overload
- No speculative data (always cite sources)
You don't even have to write it yourself. Tell Claude: "Interview me about my work style and create a CLAUDE.md."
WHY — Why does this matter?
Claude forgets the conversation when a session ends. CLAUDE.md is the "built-in memory" that survives across sessions — it eliminates the cost of repeating the same briefing every time.
Skills — recipes for repeated work
A skill is the method for a frequent task, frozen into a single file. Like a cooking recipe: create it once, and anyone gets the same quality every time.
- Using one: call it with a slash and its name, e.g.
/meeting-notes - Making one: have Claude do a task once, then say "turn what you just did into a skill so I can reuse it exactly the same way" — that's it
- Importing one: you can also install proven skills made by others. There's an ecosystem where you browse and install, like an app store
Instructions that used to take a long explanation every time shrink to a single slash word.
WHY — Why does this matter?
If CLAUDE.md is "the rules that always apply," a skill is "the procedure for one specific task." When you discover a way of working that works, you freeze that day's quality into a file — so next time is as good as today.
MCP — the USB port for external tools
MCP is the standard for connecting Claude to external services like Notion, Slack, and Google Sheets. Think of a USB port: as long as the plug fits, a keyboard or a mouse both work — one interface connects many services.
Claude Code ←→ MCP server (interpreter) ←→ External service (Notion, Slack, etc.)
Connect once, and instructions like "create a meeting-notes page in Notion" or "summarize this week's important Slack messages" become possible. The token issued during setup is best understood as "a keycard you use instead of your password."
WHY — Why does this matter?
It extends Claude's reach beyond your folder — into Notion, Slack, spreadsheets. Without MCP you'd ferry everything by copy-paste; with it, the same work becomes a one-sentence instruction.
Subagents — dividing work like a team
The main Claude acts as team lead, holding the big picture, while specialized teammates (agents) take on the detailed work.
- Role assignment: have one deliverable reviewed separately from a lawyer's perspective and an accountant's perspective
- Parallel execution: run several independent tasks at once to save time
- Verification: create a "demanding customer" agent to attack your proposal and then patch the holes it finds
WHY — Why does this matter?
One person doing everything means narrow vision and long hours. Splitting perspectives raises review quality; splitting tasks cuts time. It's the mechanism that lets even a one-person business work like a team.
Usage concepts worth knowing
| Concept | Summary |
|---|---|
| Session | One bundle of conversation. New session = fresh start |
| Plan Mode | Plan first without changing anything. Use before big tasks. Toggle with Shift+Tab |
| /compact | Compress a long conversation to recover speed and accuracy |
| /resume | Pick up yesterday's conversation today |
TIP — When to reach for Plan Mode
Before anything hard to undo, or any task likely to take 30+ minutes, get a plan first in Plan Mode. Reading the plan and correcting course before execution is far cheaper than redoing finished work.
Recap
CHECK — Part 2 in four points
- Claude Code is a harness that runs an agent safely — a tool for general computer automation, not just coding
- CLAUDE.md is the work manual, skills are recipes, MCP is the USB port, subagents are teammates — these four make it "your assistant"
- Default to Sonnet; switch to Opus for complex work
- Plan Mode before big tasks, /compact when conversations get long, /resume for yesterday's work
Now that you know the building blocks, what remains is how to delegate well. The next post covers the craft of specific instructions, verification habits, security rules, and two real-world workflow examples.